Machine learning-based EDFA Gain Model Generalizable to Multiple Physical Devices

Abstract

We report a neural-network based erbium-doped fiber amplifier (EDFA) gain model built from experimental measurements. The model shows low gain-prediction error for both the same device used for training (MSE $\le$ 0.04 dB${}^2$) and different physical units of the same make (generalization MSE $\le$ 0.06 dB${}^2$).